Use experimental values when these are available. For calculated values, the chemical properties plugin in ChemAxon MarvinSketch is considered rather accurate.

I'm gonna second ChemAxon MarvinSketch and add ALOGPS (http://www.vcclab.org/lab/alogps/). For my sets both gave very reasonable results. The one problem it had were molecules with multiple -OH groups, for which the errors with respect to experimental values were relatively high.

Thanks for your answers. I'm going to check ChemAxon calculator!

You can estimate logP values accurately using Calculation of Physicochemical properties at www.chemosophia.com . Now free and fast

www.chemosophia.com

chemoffice, cdk, rdkit

Another highly accurate commercial program for physical-chemical properties of compounds is ADMET_Predictor by SimulationsPlus.

I do not know how accurate it is, but it does not require registration and yields some other useful data such as 3D structures, polar surface area etc.: https://www.molinspiration.com/cgi-bin/properties

It is difficult to know how accurate it is. Hyperchem, ADME, Gaussian software is good for calculating LogP.

Molinspiration Cheminformatics is also a free Software available online for calculation of physicochemical properties including log P calculation.

I have used a web server for predicting pharmacokinetic properties of small-molecule using graph-based signatures.

http://biosig.unimelb.edu.au/pkcsm